Reviewer_1
Comment_1: How the initial parameters (specifically the depth of the diamond) should be selected without a priori knowledge. 
==>As Reviewer_2 mentioned, our idea is simple, but theoretical analysis is not easy. The optimal value of depth is too difficult to derive. In fact, the recommended parameters of prior sketches are also not derived from the theoretical analysis but the empirical experiments. Therefore, we conducted extensive experiments to present the empirical optimal value of depth. Our used datasets follow the Zipfian distribution, which is a common distribution in practical datasets according to literatures [10, 11]. Therefore, our recommended parameters are suitable for many real data streams. 

Comment_2: What is the impact of getting these "wrong"?
==>In this case, Diamond will not reach its highest performance, but it will still outperform prior art with high probability. We will add relevant experimental results in our paper.

Comment_3: If they were to use more memory (10MB), would they be able to reduce the depth of the Diamond and therefore computation complexity? What impact would this have on throughput and accuracy? 
==>Indeed, when using 10MB, we can reduce the depth of the Diamond to 2 to achieve faster query and update speed, while keeping the accuracy almost unchanged. We will add these experimental results in our paper.

Comment_4: Why did the authors choose the BOB-Hash? What made this better than other hash functions?
==>The code of BOB-Hash is open sourced. It meets our demands, for we can choose different seeds and get different sets of independent hash functions. We tried using other hash functions, and the accuracy of all sketches keeps almost unchanged.  
We have provided interfaces in our source codes which are available at GitHub for users to choose different hash functions. 

Comment_5: Section 3.1 seems unnecessaryˇ­
==>We will remove it.

Reviewer_2:
Comment_1: I am not sure if SC is the correct venue for the manuscript. There is no parallelization or any arguments for the distributed stream setting, e.g., to combine multiple sketches. CM and some other sketches obtained from multiple streams can be combined. How about Diamond Sketch and what is the procedure for that?
==>Sketch is a compact data structure that achieves the FASTEST query and update speed for high-speed data streams, which is highly related to the topic "High Performance Networking Technology" of SC. Therefore, we insist our paper is appropriate for SC. Sketches can be queried in parallel naturally. For example, different atom sketches can be queried in parallel using different threads. 
We can also apply "batch-processing" parallelization to our Diamond sketch. 1) We conducted GPU experiments for the paralleled query of the Diamond sketch. 2) One of our undergoing work is implementing Diamond on FPGA platform. We will add these results in the camera-ready version if accepted. Inspired by the comments, we propose a new scheme to combine two Diamond sketches A (a_1 ˇ­ a_d) and B (b_1 ˇ­ b_d) by adding atom sketch a_i and b_i with an offset. Specifically, we shift the whole b_i to right by 1 (counter), and then add a_i and b_i counter by counter. When querying an item, the only difference is to read two adjacent counters for each hashing. More details will be provided in the revised paper.


Comment_2: Will a hybrid sketch work using a different existing sketch? Do the atom sketches always need to be identical? 
==>Diamond can use hybrid atom sketches. We will add more discussions about it in our paper.


Reviewer_3
Comment_1: It is unclear how many real life transactional datasets have a non-uniform distribution. 
==>Many non-uniform datasets follow the Zipfian distribution. More details are provided in the response to Comment_1 of Reviewer_1.

Comment_2: The parameter choiceˇ­how well it will influence the results.
==>Please refer to the response to Comment_1 and 2 of Reviewer_1.

Comment_3: The throughput achieved by the new method is sometimes half of related work. A larger discussion related to the expected scalabilityˇ­would be good.
==>The throughput of Diamond during insertion and query is comparable to other sketches except the CSM sketch, which achieves the highest speed at the cost of an intolerable drop in accuracy. The throughput of our Diamond sketch can achieve 4~6 Mops, which can meet the requirements in a broad range of practical scenarios. We would add more discussions of scalability of our method in our paper. 

Comment_4: What is the memory consumption of CSM, CM, A, C, CU sketches and how they scale in the future?
==>As mentioned in 5.2.2, unless explicitly stated, the memory consumption of all sketches is 0.5MB.